Python Plays Grand Theft Auto 5 - Reboot

This is a reboot of the project from 2017, but with a whole new approach. This time, instead of collecting data by hand and training a classification model, we opted to create a whole new system for the data collecting, training and playing, and to do all of that live. Our models are now using regression (with other output types being still possible), training data is collected automatically and we stream models during their training phase so anyone can join and watch the progress.

Our AI, besides having multiple alter-egos (models), is still called Charles, like in 2017.

For a full explanation, refer to the System page, but for a quick summary - we’re using a central server that all other parts connect to. The Data Collectors run in separate GTA5 instances and our NPCs are collecting and Balancing the data, which is sent through the Server to the Trainer. The Trainer buffers these data, trains the model and updates the Player. The Player is using the Dual-Camera system to play so we can watch the 3rd person camera (also called cinematic camera) while the model is fed the Hood Camera.

For more information, refer to these other pages: - this page - main page with the project progress - System - describing how all parts are working together to train and inference the models - Cameras - how we created multiple cameras - Convcam - what the Convcam is and how we use it - Data Balancing - how we are balancing regression data - NPCs - our custom NPCs playing in the GTA5 to collect data automatically - Purpose - a way to let the model “know” where to drive - Storage and Buffer - how are we managing the training data and why random batches are important - Unstuck - how to make the car not be stuck anywhere - Xception - the first CNN backbone architecture used - InceptionResNetv2 - the second, more successful, CNN backbone architecture used - Tensorboard logs - all of our Tensorboard logs for all of the trained models - The Original Project From 2017 - model_0001_xception - the first logged Xception model, Balancing_v1 - model_0002_xception - trying Balancing_v2 and a custom model - model_0003_xception - Balancing_v3 and back to a standard model - model_0004_inceptionresnetv2 - trying a bigger model - InceptionResNetv2 - model_0005_inceptionresnetv2 - transfer-learning attempt - model_0006_inceptionresnetv2 - added stacked images - model_0007_inceptionresnetv2 - trying a history inputs - model_0008_irv2_data_td - history with recurrent layers - model_0009_irv2_cr_tl - history with recurrent layers with transfer-learning - model_0010_irv2_tcb - dual-backbone InceptionResNetv2 - model_0011_x_tcb - dual-backbone Xception - model_0012_regnet - TBA

Stream and current stream layout:

The progress is being streamed on Twitch: https://www.twitch.tv/sentdex. Below we keep a list of dates and streamed models.

Stream's layout consists of:
- top-left tile is the 3rd person camera of the car driving, the main camera to observe Charles - this main tile, in its lower-right corner, shows the current model and when it's been created - top-right tile called Hood Camera is exactly what the models "see" - the input to the convolutional backbone of the model - middle-right tile called Convcam shows the reshaped output of the CNN backbone and lets us observe how the CNN part trains - bottom-right tile called Player Console shows current driving predictions along with additional information - bottom-middle tile called Server/Trainer Console shows training progress along with some basic training information - bottom-left tile called Tensorboard shows the loss of the training process

Driving examples:

Model list:

• The Original Project From 2017 - how this started
• model_0001_xception - the first logged Xception model, Balancing_v1
• model_0002_xception - trying Balancing_v2 and a custom model
• model_0003_xception - Balancing_v3 and back to a standard model
• model_0004_inceptionresnetv2 - trying a bigger model - InceptionResNetv2
• model_0005_inceptionresnetv2 - transfer-learning attempt
• model_0006_inceptionresnetv2 - added stacked images
• model_0007_inceptionresnetv2 - trying a history inputs
• model_0008_irv2_data_td - history with recurrent layers
• model_0009_irv2_cr_tl - history with recurrent layers with transfer-learning
• model_0010_irv2_tcb - dual-backbone InceptionResNetv2
• model_0011_x_tcb - dual-backbone Xception
• model_0012_regnet - TBA

Streaming timeline

May 1st: - Started streaming model_0003_xception_v2 from scratch

May 2nd: - Stopped streaming model_0003_xception_v2 at batch 15250 - Started streaming model_0003_xception_v1 from batch 24500 (pre-trained off-stream)

May 7th: - Stopped streaming model_0003_xception_v1 at batch 162650 - Started streaming model_0004_inceptionresnetv2_v1 from scratch - Stopped streaming model_0004_inceptionresnetv2_v1 at batch 1250 - Started streaming model_0004_inceptionresnetv2_v2 from scratch - Stopped streaming model_0004_inceptionresnetv2_v2 after ~160 batches - Started streaming model_0003_xception_v1 from batch 162650 (continuing training) - Stopped streaming model_0003_xception_v1 at batch 165250 - Started streaming model_0004_inceptionresnetv2_v1 from batch 1250 (continuing training)

May 8th: - Stopped streaming model_0004_inceptionresnetv2_v1 at batch 12000 - Started streaming model_0005_inceptionresnetv2_v1 from scratch - Stopped streaming model_0005_inceptionresnetv2_v1 at batch 3500 - Started streaming model_0004_inceptionresnetv2_v3 from batch 22250 (pre-trained off-stream)

May 14th: - Stopped streaming model_0004_inceptionresnetv2_v3 at batch 184250 - Started streaming model_0006_inceptionresnetv2_v1 from batch 18750 (pre-trained off-stream)

May 16th: - Stopped streaming model_0006_inceptionresnetv2_v1 at batch 60500 - The stream has been offline for several days since we had to send back the Comino machine and setting up another one took longer than we expected

May 28th: - After new machine was set and new models developed, we restarted streaming - Started streaming model_0009_irv2_cr_tl_v1 from scratch

May 29th: - Stopped streaming model_0009_irv2_cr_tl_v1 at batch 19250 - Started streaming model_0004_inceptionresnetv2_v3 from batch 184250 (continuing training)

July 5th: - Stopped streaming model_0004_inceptionresnetv2_v3 at batch 752000 - The stream has been stopped as there have been no further improvements in the model and to catch up with documentation and do some off-stream development

Project timeline

This is not a full project log and contains only the key changes that are visible on the stream (and a few from before the stream has started). There are many more things going on "behind the scenes".

2017: - Development of The Original Project From 2017

November 2021: - The idea of bringing back this project has born

Early 2022: - The project has started the development phase of the main system, the initial models, and the game mod

April 8th: - Finished Balancing_v1 - Started training model_0001_xception_v1 from scratch - Stopped training model_0001_xception_v1 at batch 10800 - Started training model_0001_xception_v2 from scratch

April 9th: - Stopped training model_0001_xception_v2 at batch 7150 - Started training model_0001_xception_v3 from scratch - Started training model_0001_xception_v4 from scratch - Stopped training model_0001_xception_v3 at batch 18850 - Started training model_0001_xception_v5 from scratch

April 10th: - Stopped training model_0001_xception_v5 at batch 25200 - Stopped training model_0001_xception_v3 at batch 23300 - Started training model_0001_xception_v6 from scratch

April 11th: - The first data histogram has been created to see the Data Balancing

April 15th: - Stopped training model_0001_xception_v6 at batch 176850

April 19th: - Started training model_0002_xception_v1 from scratch

April 22nd: - Started training model_0003_xception_v1 from scratch

April 24th: - Stopped training model_0003_xception_v1 at batch 24500

April 30th: - Stopped training model_0002_xception_v1 at batch 249750

May 1st: - The Convcam has been created - Started streaming on Twitch.tv - Started streaming model_0003_xception_v2 from scratch

May 2nd: - Stopped streaming model_0003_xception_v2 at batch 15280 - Started streaming model_0003_xception_v1 from batch 24500 (pre-trained off-stream) - Started training model_0003_xception_v3 from scratch

May 3rd: - Stopped training model_0003_xception_v3 at batch 17500 - Started training model_0003_xception_v4 from scratch - Stopped training model_0003_xception_v4 at batch 5150

May 7th: - Stopped streaming model_0003_xception_v1 at batch 162650 - Started streaming model_0004_inceptionresnetv2_v1 from scratch - Stopped streaming model_0004_inceptionresnetv2_v1 at batch 1250 - Started streaming model_0004_inceptionresnetv2_v2 from scratch - Stopped streaming model_0004_inceptionresnetv2_v2 after ~160 batches - Started streaming model_0003_xception_v1 from batch 162650 (continuing training) - Stopped streaming model_0003_xception_v1 at batch 165350 - Started training model_0004_inceptionresnetv2_v3 from scratch - Started streaming model_0004_inceptionresnetv2_v1 from batch 1250 (continuing training)

May 8th: - Stopped streaming model_0004_inceptionresnetv2_v1 at batch 12000 - Started streaming model_0005_inceptionresnetv2_v1 from scratch - Stopped streaming model_0005_inceptionresnetv2_v1 a